CNN_Binary_Image_Classification

Goal

Design and train a Small Convolutional Neural Network in Julia and FluxML that accurately classifies images into either 2 classes:

• Dog (1)
• Cat (0)

Design

Neural Network Architecture

• Convolutional Layers: 3
  • Input Feature Maps: 3
  • Ouput Feature Maps: 16
  • Activation: ReLU
• Pooling Layers: 3
  • Kernel Size: 2x2
  • Position: Directly after each Convolutional Layer
• Flattening Layer: 1
  • Position: After last Pooling Layer
• Dense Layers: 3
  • $1^{st}$ nodes: 16384
    • Activation: ReLU
  • $2^{nd}$ nodes: 5250
    • Activation: Sigmoid
  • $3^{rd}$ nodes: 1

Hyperparameters

• Learning Rate ($\alpha$): 0.01
• Momentum ($\psi$): 0.0001
• Kernel Size ($\kappa$): 3x3
• Stride ($\zeta$): 1
• Padding ($\rho$): 0
• Weight Decay ($\lambda$): 0.0004
• Batch Size: 128

Training

• Loss Function: Log Cross Entropy
• Optimizer: Gradient Descent ($\alpha$ , $\psi$)

Training Metrics

Loss Log

At Last Epoch

Training Loss: 0.1

Validation Loss: 0.1

Accuracy Log

At Last Epoch

Training Accuracy: 97.07

Validation Accuracy: 97.1

Test

Test Accuracy: 97.15%